A website can comply with technical accessibility rules and still be practically impossible to use. What good is it for the code to look correct if someone navigating with a keyboard cannot tell where the focus is or cannot complete a form?
That is the problem Brendan Works, a technical product manager specializing in adapted transportation services, wanted to address with PointCheck. The tool combines browser automation and open-source AI models to evaluate how a page behaves on screen, not just how it is written internally.
The limits of code-based testing
Most accessibility tools inspect a site's HTML, styles, and technical structure. They then compare those elements with established standards. That process is valuable, but it has a blind spot: code can comply with a rule and still produce a confusing or unusable experience.
Think about the focus indicator. When you navigate with the Tab key, a visible outline should appear around the selected link or button. For someone who uses only a keyboard, that outline serves a purpose similar to the mouse cursor: it shows where the next action will happen.
But a site can define that indicator correctly in a stylesheet and make it practically invisible because its color blends into the background. A test that only analyzes the code might approve it. A user, on the other hand, would not know whether they are about to press "Buy now" or "Cancel."
Accessibility is not just about valid code. It also means that the interface is perceivable, operable, and understandable to the person using it.
How PointCheck works
PointCheck uses Playwright, an open browser automation framework. The application loads a page in a Chromium window in the background and simulates the navigation of someone using a keyboard.
Before it begins, a script looks for common problems, such as elements that trap focus, broken links, poorly implemented interactive controls, and tabindex values that alter the natural navigation order.
Next, PointCheck moves through the page with the Tab key and takes a screenshot at each step. This is where the Ai2 models come in:
- MolmoWeb identifies the location of the focused element in the image using pixel coordinates.
- Molmo analyzes that region to determine what appears on screen and whether the focus indicator is actually visible.
- Olmo 3 turns the technical results into a report written in clear language.
The difference matters. Instead of asking only whether a CSS rule related to focus exists, PointCheck tries to verify whether a person can see that focus in the rendered interface.
More than a visual test
The tool also includes tests at 200% zoom, contrast checks, and a simulation of deuteranopia, one of the most common forms of altered color perception.
It also evaluates forms with invalid data, reviews the markup structure, and analyzes keyboard navigation issues. These tests run as separate modules because each one looks for different kinds of failures.
Olmo 3's final report is designed for designers, product managers, and development teams, not just accessibility specialists. That translation matters: detecting a problem is only the first step. Understanding it and knowing how to prioritize it are also part of the work.
Open models for protecting sensitive data
Works evaluated using proprietary vision and language APIs, but chose an architecture that can be hosted independently. With Molmo and Olmo, teams can download the models' weights, run them on their own infrastructure, and keep screenshots of their sites from leaving their environment.
For a company working with medical, financial, or government information, that difference can be decisive. Screenshots do not always contain public data. A self-hosted approach makes it possible to maintain greater control over privacy, processing, and operating costs.
The first PointCheck prototype was built in a single weekend. Automating the process with Playwright was relatively straightforward. The more complex technical challenge was running all three models on a single NVIDIA A10G GPU without exhausting video memory, while also resolving compatibility issues in the inference code.
That detail highlights a reality that is often lost in AI demonstrations: having open models does not eliminate the engineering work. You still have to manage memory, dependencies, latency, and deployment. The advantage is that the team can inspect and adapt every component instead of depending on a black box.
A support tool, not a replacement for human evaluation
PointCheck is not meant to replace traditional tools or validation with users. Its goal is to cover an area that purely structural tests cannot observe: what people actually see and what happens on the screen.
Technology-assisted navigation depends on details that may seem minor to other users. An invisible focus indicator, a strange tab order, or a button that does not respond to the keyboard can prevent someone from booking transportation, attending classes, getting to a medical appointment, or maintaining their social life.
The proposal also reveals one of the strengths of open models: enabling independent developers to build solutions for specific needs, with control over data and infrastructure. How many useful tools still do not exist because their creators cannot access models they are free to run and adapt?
Web accessibility is not some future scenario or an optional feature. It is the difference between a platform that offers an opportunity and one that leaves some of its users out. PointCheck adds a visual and practical layer to that process, while showing that open AI can become concrete software for improving everyone's digital experience.
Original source
https://allenai.org/blog/global-accessibility-awareness-day-2026